Why Walk-Forward Backtesting Is the Only Backtest That Matters
2026-02-18 · endeavr.ai Research
If you have ever seen a trading strategy with a gorgeous equity curve that fell apart the moment real money touched it, you have met the most expensive bug in quantitative finance: a backtest that lied. The usual culprit is the validation method itself.
The problem with a single split
The naive approach trains a model on most of history and tests it on a held-out chunk. The trouble is that financial markets are non-stationary — the relationship between features and returns drifts as regimes change. A model tuned on one static split can be silently overfit to the conditions of that period, and a single hold-out number tells you nothing about how it degrades over time.
How walk-forward testing works
Walk-forward validation trains on a rolling window, predicts the next out-of-sample period, then slides the window forward and repeats. Every prediction uses only information available at that moment. The output is not a single accuracy figure but a realistic, time-ordered track record you could actually have traded.
What endeavr.ai does
Every model in our four-architecture ensemble is walk-forward backtested nightly before any weights reach production. We publish directional accuracy, Sharpe, and drawdown versus buy-and-hold per ticker. If a model degrades, you see it — we do not hide underperformance behind a cherry-picked window.